nengwp/cellgan

[MICCAI 2023] Official Pytorch Implementation for "CellGAN: Conditional Cervical Cell Synthesis for Augmenting Cytopathological Image Classification"

CellGAN-for-Cervical-Cell-Synthesis

Official Pytorch Implementation for "CellGAN: Conditional Cervical Cell Synthesis for Augmenting Cytopathological Image Classification" (Early Accepted in MICCAI 2023 https://link.springer.com/chapter/10.1007/978-3-031-43987-2_47)

Method

CellGAN synthesizes 256×256 cytopathological images of different cervical squamous cell types (NILM, ASC-US, LSIL, ASC-H, and HSIL). It can serve as a data augmentation tool for patch-level cell classification in automatic cervical abnormality screening.

Qualitative Results

Environment

• Python 3.10.10
• Pytorch 2.0.0+cu117
• opencv-python, scikit-image, tqdm

Quick Start

1. We provide a pre-trained CellGAN generator checkpoints/model.pth for synthesizing cytopathological images.

2. Use the following command to synthesize a certain number of images for a desired cervical cell type.

python cellgan_inference.py --config [config_name] --model [model_path] --output_dir [directory to save generated images] --cell_type [desired cell type] --data_num [number of generated images]

Usage

Training

• Refer to configs/default_config.yaml for customizing your own configuration file configs/{config_name}.yaml. All the arguments are self-explanatory by their names and comments.

• Set the argument DATAROOT in configs/{config_name}.yaml to your training data root.

• In DATAROOT, split your images into subdirectories according to the cell types and prepare an img_list.txt.

• The directory structure of DATAROOT should be prepared as in the following example:

DATAROOT
├─ NILM
|  ├─ NILM_image_0001.png
|  └─ ......
├─ ASC_US
|  ├─ ASC_US_image_0001.png
|  └─ ......
├─ LSIL
|  ├─ LSIL_image_0001.png
|  └─ ......
├─ ASC_H
|  ├─ ASC_H_image_0001.png
|  └─ ......
├─ HSIL
|  ├─ HSIL_image_0001.png
|  └─ ......
└─ img_list.txt

• The TXT file img_list.txt should contain one image path {category_name}/{image_name}.png per line as in the following example.

NILM/NILM_image_0001.png
NILM/NILM_image_0002.png
......
ASC_US/ASC_US_image_0001.png
......

• After finishing data preparation, use the following command:

python train.py --config [config_name]

Testing

Edit the testing arguments in configs/{config_name}.yaml and use the following command:

python test.py --config [config_name]

Literature Information

Authors:

Zhenrong Shen[1], Maosong Cao[2], Sheng Wang[1,3], Lichi Zhang[1], Qian Wang[2]*

Institution:

[1] School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China

[2] School of Biomedical Engineering, ShanghaiTech University, Shanghai, China

[3] Shanghai United Imaging Intelligence Co., Ltd., Shanghai, China

Manuscript Link:

https://arxiv.org/abs/2307.06182 (preprint on arXiv)

https://link.springer.com/chapter/10.1007/978-3-031-43987-2_47 (MICCAI 2023, conference version)

Citation:

@inproceedings{shen2023cellgan,
  title={CellGAN: Conditional Cervical Cell Synthesis for Augmenting Cytopathological Image Classification},
  author={Shen, Zhenrong and Cao, Maosong and Wang, Sheng and Zhang, Lichi and Wang, Qian},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={487--496},
  year={2023},
  organization={Springer}
}